| name | seedance-2-prompt-standard |
| description | Expert-tier prompting protocol for Seedance 2 (ByteDance) video generation. Focuses on director-style temporal segmentation, physical descriptors, and cinematic metadata to prevent 'robotic' defaults. |
| version | 1 |
Seedance 2 Prompting Standard
Seedance 2 rewards precise, structured, film-director language over loose descriptions. It excels at physics and multi-shot coherence when provided with temporal segmentation and physical/lighting/camera specs.
Core Principles
- Physics > Vibes: Use specific physical descriptors (e.g., "hydraulic hiss", "metal scrape", "steam bursts") rather than vague adjectives.
- Camera Language: Use specific cinematography terms (e.g., "ultra-slow push-in", "side tracking shot", "low-angle from ground level").
- Lighting: Use physical, source-based descriptors only (e.g., "soft morning light through a window on the left") instead of "cinematic lighting".
- Skin/Texture: Avoid waxy defaults by specifying "realistic skin texture, visible pores, natural slight unevenness".
- Motion vs. Camera: To avoid artifacts, prefer either a static camera with a moving subject, or a controlled camera move. Simultaneous high-speed movement in both can cause confusion.
- Image-to-Video Rule: Describe only motion and camera; never re-describe the reference image. Lead with Subject + Primary Action in the first 20-30 words.
- Compiler Requirement: Every prompt must include temporal segmentation or explicit shot progression, one camera behavior, one source-based lighting phrase, one physical texture phrase, and one dedicated negative line.
Prompting Formats
1. Timed Paragraph (Standard Narrative)
Best for: General cinematic clips, character acting, and narrative sequences.
Structure:
[CINEMATIC SETUP]
[Film stock / lens / aperture / camera behavior]. [Color grade]. [Lighting source]. [Atmosphere]. [Audio: sound FX only / no music]. Face stable, no deformation.
[@ image1] is [role/description]. [@ image2] is [reference for X].
0-2s: [Shot type]. [@ image1] [precise action]. [Camera move]. [Sound.]
2-3s: [Shot type]. [Action sequence with physics]. [Camera behavior]. [Sound FX.]
...
[Final timestamp]s: [Closing shot]. [Final action / movement]. [Sound fade or impact.]
Negative Line (Append to each sequence):
negative: no face morph, no costume disappearing, no armor floating, no floating limbs (Adjust based on specific risk).
2. JSON Format (Complex VFX/Physics)
Best for: High-precision control over complex effects like POV time-freezes or intricate mechanical movements.
{
"shot": { "composition": "...", "lens": "...", "camera_movement": "..." },
"subject": { "description": "...", "wardrobe": "...", "props": "..." },
"scene": { "location": "...", "environment": "..." },
"visual_details": { "action": "...", "special_effects": "..." },
"cinematography": { "lighting": "...", "color_palette": "...", "tone": "..." },
"audio": { "music": "...", "sound_effects": "..." }
}
Failure Modes to Avoid (The "Kill List")
- Vague Descriptions: "A woman holding a product" $\rightarrow$ Robotic. Use full scene + lighting + camera + emotion.
- Waxy Skin: Missing texture descriptors.
- Polished Fake Look: Using "professional", "high quality", or "studio". Counter with "handheld iPhone", "casual slightly unsteady framing".
- Smooth AI Look: Missing pores, cloth weave, flyaway hair, sensor noise, sweat, dust, scratches, or material wear.
- Neutral Default: Unspecified emotion. Always define the emotional state (e.g., "genuine surprise building to pleased recognition").
- Vague Lighting: "Cinematic" or "black" without specific source/intensity.
- Moving Camera + Moving Subject Simultaneously: Causes artifacts and motion confusion. Use one or the other; static camera while subject moves is usually the safer choice.
Debated Techniques (Community Split)
"Cinematic" / "4K" / "IMAX quality" suffix tags: Widely used as a polish line at the end of prompts (e.g., cinematic, 4K, film quality). Some power users report a visible quality boost; others argue the gain is marginal and that strong physical descriptors earlier in the prompt matter more. Both camps continue using them. Safe to include as a final line — unlikely to hurt, may help.
Vocabulary Cheat Sheet
- Push-in: "ultra-slow push-in", "almost imperceptible push-in".
- Tracking: "side tracking shot moving with [subject]", "ground-level tracking".
- Angles: "extreme macro", "low angle from floor", "POV time-freeze", "forced-foreground low-angle".
- Feel: "handheld phone camera feel with slight natural movement", "locked-off feel. No handheld".
- Slow-mo: Use "240fps feel" or "half-speed" instead of just "slow motion".